AI Tutors Are Reaching Classrooms. What Are the New Trials Showing?
A student makes a mistake in a mathematics problem. An AI tutor can offer a hint immediately, but what happens next matters: does the student reconsider the idea, ask a useful question, or simply move toward the next answer? New experimental research is making that question more concrete.
Schools considering AI tutoring now have more to examine than demonstrations and enthusiastic predictions. Recent working papers investigate how students actually use these systems and what happens to learning. These are working papers, so their findings remain open to revision and further scrutiny. They invite careful decisions about the task, the support around it, and the evidence used to judge success.
Similar technology, different learning conditions
An August 2026 NBER working paper on Khanmigo reports a two-year experiment across 18 Tennessee middle schools. Students used Khan Academy with an AI tutor during remedial mathematics sessions. Achievement gains resembled those from Khan Academy practice without AI assistance, and substantive dialogue with the tutor was uncommon. Access to a conversational tool did not ensure that students used conversation to think through mathematics.
A separate August working paper studied more than 6,000 middle-school students using an AI-supported mathematics practice platform. The researchers found that the organization of practice mattered. The most encouraging delayed-test results appeared when AI support was embedded within a mastery workflow, with limited evidence concentrated on particular practiced material.
An October 2026 university trial adds a different perspective. In that setting, access to a course-integrated tutor was associated with reduced participation in existing learning activities and lower grades in the within-course comparison. These studies examine different learners, tools, tasks, and comparisons. They do not establish a single verdict on AI tutoring.
Ask what the tutor changes
A useful school pilot begins with a specific learning difficulty. Perhaps students repeatedly confuse two fraction concepts, or stop working when the first attempt is wrong. The team can then ask what an AI tutor would help students do that the current approach does not support well.
For example, a teacher could ask students to explain why a denominator stays the same when adding fractions with a shared denominator. A useful hint might direct attention to the size of the units. A less useful interaction could supply a completed solution before the learner has considered the relationship. Both interactions may look productive on a dashboard, but they create different opportunities to learn.
Check what students can do afterward
Teachers need to see more than the number of questions completed. They can examine the reasoning students produce, the kinds of help requested, and what learners can explain later without the tool. A short follow-up task using a different example can reveal information that an immediate correct answer misses.
Participation deserves attention too. If students spend less time discussing, reading, or practising because a tutor is available, a pilot should examine that change. Time saved is useful only when the new pattern serves the learning goal.
Make adoption a learning process
Before a pilot, schools can agree on its purpose, duration, accessibility arrangements, privacy requirements, and signs that would justify continuing or changing course. Families should understand what the tool does and when students are expected to work independently.
The next step in AI tutoring is an opportunity to become more precise about support. What would students need to understand better for a school to say that its AI tutor was helping?
Illustrative image generated with AI.
